Supervised Learning Function for Neural data

This toolbox facilitates neural decoding of time series. Under the hood it uses scikit-learn functions.

Set-up

Install using pip (https://pypi.org/project/temp-dec/)

pip install temp_dec

Requirements

sklearn==0.21.3
numpy==1.13.1
scipy==0.19.1

Python implementation

This function takes in the data X (ndarray; trials by features by time), labels y (ndarray; vector).

from temp_dec import decoding_functions
decoding_functions.temporal_decoding(X, y)

Using a sliding time window

If there is information in the temporal dynamics of the signal, using a sliding time window will increase decoding accuracy (and smooth the signal). We can also demean the signal within each window, this avoids the issue of baselining.

size_window=5
demean=True

Applying PCA

If you use a large amount of features, you might want to consider applying PCA to your features before applying your classifier. In addition, classifiers are sensitive to noise rejecting noise components from the data can be beneficial. You can also regulate how many components you would like to keep (setting the pca_components variant to > 1) or how much variance you would like to explain (setting the pca_components variant to < 1). As a general rule of thumb maintaining 95% of variance will maintain enough signal and reduces feature space. If pca_components = 1 then 100% of the variance will be maintained so no PCA is applied.

pca_components = .95

Classifiers

Different classifiers are supported, selected in accordance with Grootwagers et al (2017) j.cogn.neurosci.

• LDA: linear disciminant analysis
• LG: logistic regression
• GNB: Gaussian Naive Bayes
• maha: Nearest Neighbours using mahalanobis distance.

classifier = 'LDA'

with the amounts of stratified cross-validations (kfold) adjusted with the following flag, 5-fold by default.

n_folds = 5

All options incorporated

output = decoding_functions.temporal_decoding(data,labels,
                                                n_folds=5,
                                                classifier='LDA',
                                                pca_components=.95,
                                                size_window=20,
                                                demean=True)